Causal Inference

The main goal of this project is to build libraries and models rooted in deep learning to estimate the causal effects of an intervention on some measurable outcomes (e.g. the effect of a treatment or procedure on survival).

Installing dependencies

pip install --upgrade pip
pip install -r requirements.txt

Performing estimation

While in the base directory of the repo, run:

python estimate.py

The results of all estimators will be stored in folder estimation_results.

Available estimators:

• COM/S-Learner: OLS1, RF1, NN1;
• GCOM/T-Learner: OLS2, RF2, NN2;
• Inverse Probability Weighting (IPW);
• Double Machine Learning/R-Learner;
• TARNet
• Dragonnet

where OLS stands for Ordinary Least Squares, RF stands for Random Forest, and NN stands for Nerual Networks. They represent ML models that are linear, ensembled, and non-linear respectively.

Available datasets:

Please see the description of the datasets in the data folder.

• IHDP-100
• Jobs
• TWINS

Available metrics:

• Mean absolute error (MAE) of ATE
• Precision in Estimation of Heterogeneous Effect (PEHE)

Performing nusiance model estimation

The nuisance models are the models used in the paper Empirical Analysis of Model Selection for Heterogenous Causal Effect Estimation to estimate the ground truth for PEHE calculation.

While in the base directory of the repo, run:

python utils/nuisance_model_selection.py --dataset <dataset_name>

This uses AutoML to estimate the best model for each nuisance model. The results are stored in the estimation_results/<dataset_name>/models folder. The results for Jobs and TWINS are already stored in the repo.

Available estimators:

• T-Learner 0
• T-Learner 1
• S-Learner
• Double Machine Learning (DML)
• Propensity

Available datasets:

• IHDP-100
• Jobs
• TWINS

Performing evaluation

After performing estimation, run:

python evaluate.py

Running Notebooks

• Open the notebooks folder and click on the desired notebook.
• At the top of the notebook, click on the button.
• Follow the instructions in the Colab to run the code.

For the Demo_End2End_Causal_Estimation_Pipeline.ipynb notebook, to ensure that you've initialized it correctly, check the following:

• After running the first code block in Colab, your file directory should look like this:

(Note: you can press to refresh)

using pre-commit hooks

To check your code at commit time

pre-commit install

You can also get pre-commit to fix your code

pre-commit run